Fish schooling evolved about 60 times and was lost about 320, study finds

Machine learning scored collective movement in 2,839 ray-finned fish species, nearly a tenth of those alive today, and the resulting family tree shows the habit was lost far more often than it was gained.

Peer reviewed · Comparative (observational) study · Traits inferred by machine learning

Moving as a coordinated group has appeared in ray-finned fishes about 60 separate times and disappeared about 320 times, according to a study by Jay Love, Andrew Hein and colleagues at Cornell University, the University of Colorado Boulder and Florida Atlantic University, published in PNAS on 6 November 2026. Schooling tracked heavier predation, patchy food and faster growth in species numbers. Its authors call it one of the largest studies of how any behaviour has evolved. Every link it reports is a correlation, and the behaviour scores came from algorithms reading text and photos, not from watching each species.

Evidence at a glance

Is this a big deal?

It is a solid step in scale, and an early one on causes. The authors describe it as among the largest studies of behavioural evolution in any group of species, and Hein called the frequent losses the more surprising result. The ecological and diversification links are correlations from one modelling study and need testing with other methods and with field observation.

What scientists found

Ray-finned fishes, the class Actinopterygii, are the most species-rich class of vertebrates, as the paper notes. The team scored whether each species shows collective movement: individuals swimming together as a unit, lined up with their neighbours. The paper’s abstract reports “over 300 evolutionary transitions” between schooling and solitary life. Cornell’s release breaks that down as about 60 gains and about 320 losses over the last 350 million years of fish evolution.

That ratio is the headline result. SciQuest’s arithmetic: 320 ÷ 60 is about 5.3, so schooling was dropped more than five times for every time it arose. Hein told the Cornell Chronicle that the number of losses was perhaps the more surprising finding.

The trait also behaved unevenly. In the abstract’s words, collective movement is “highly evolutionarily stable in some clades and extremely labile in others”. In some branches of the tree it barely changed; in others it flipped back and forth.

How they found it

Watching thousands of species swim was never an option, so the team taught software to do the reading and looking. Cornell’s release says a text transformer, the same kind of model behind large language models, turned published studies and species descriptions into numbers that indicate whether a species likely moves collectively. A second method took citizen-science photos and videos of fish from iNaturalist and other sources, reconstructed them in three dimensions, worked out which way each fish was heading, and checked whether it was aligned with others.

Together the two methods gave schooling scores for 2,839 species, which the abstract describes as nearly 10% of living ray-finned fishes. The researchers then placed those scores on the evolutionary tree of fishes and estimated where the trait switched on and off. Those switches are inferred from the tree, since nobody can watch an extinct ancestor swim.

They also asked what else schooling species have in common. Species lower in the food chain, which face more predators, were more likely to school. “We did find this correlation exists,” Hein said, adding that the relationship isn’t iron clad: many species low in the food chain swim alone and many near the top school. Predators that move in groups, he said, virtually all eat prey that also move in groups.

Why it matters to you

This won’t change anything in your daily life. It speaks to the wonder lens: why some fish in a reef video move like one animal while their close relatives drift alone. A fair comparison is people drifting toward the busiest stall at a food market. Following others pays off when they are heading for the same food. The comparison has limits: a shopper chooses on the day, while schooling is an inherited habit that this study follows over millions of years.

What the evidence does not show

It does not show that predators made fish school, or that schooling made new species. The abstract says the evolution of collective movement “coincides with an increase in net macroevolutionary diversification rate”. Net diversification is new species forming minus old ones going extinct. Love put the open question plainly in Cornell’s release: “We don’t know if the speciation rate is higher or the extinction rate is lower, but we know that through a combination of those two things, we have more species being generated.”

Linking a trait to diversification is a known trap. A 2015 analysis in Systematic Biology by Daniel Rabosky and Emma Goldberg found that a widely used family of models flagged traits simulated with no effect at all as significantly linked to speciation, with “disconcerting ease”. Part of that problem came from traits that change only rarely, and schooling changed hundreds of times, which helps. SciQuest could not read the full paper, so we cannot say which models or checks the authors used against that risk.

The trait scores carry their own uncertainty. A species counts as collective because text and images say so, and some species may be misjudged where descriptions are thin or photos are few. The paper’s error rates for those calls were not available to us.

How this fits what we already knew

Evolutionary biologists usually think of traits as evolving through selection on each individual’s own features, as the abstract notes. Collective movement is different: it only exists when individuals interact, yet it can decide whether each one survives. Cornell’s release describes two leading explanations for schooling: safety in numbers, and help finding food. Love said in the release that the predator-avoidance idea has “never been broadly tested” across species. This study is that broad test, and both ideas fit its correlations.

The many losses point to costs. Cornell’s release lists three that the team suspects: fish in groups are more open to misinformation, sharing scarce food can make grouping a poor deal, and dense groups spread disease. Hein, Gil and Fahimipour were among the authors of a 2023 PNAS study that tracked false alarms spreading through groups of wild coral reef fish.

Independent view

SciQuest found no on-record comment from an independent researcher about this paper as of 9 November 2026 and was unable to request one during this automated run. We will add one if we obtain it.

How other headlines framed it

No independent outlet had covered the paper when we checked, so the useful comparison is between the study’s own headlines. The bioRxiv preprint in April said collective behaviour “drives” diversification, a causal word. The PNAS title describes “evolutionary dynamics” without claiming a cause. Cornell’s release leads with the method, saying machine learning helps explain the evolution of fish schools. SciQuest’s headline reports the gains and losses, the part the paper measures most directly.

What happens next

The open questions are the ones the team named: whether schooling lineages gain species by splitting faster or by dying out more slowly, and which costs drive so many losses. Answering them will take other diversification methods, checks on the machine-learning scores, and field studies of what grouping costs. We will update this story if independent researchers respond. For another study that inferred animal behaviour from computation, see our report on how fly connectome simulations point to a three-neuron walking rhythm.

In one sentence

Across 2,839 ray-finned fish species, swimming in coordinated groups arose about 60 times and vanished about 320 times, and it tracks with predators, patchy food and faster species growth, though none of those links is yet shown to be a cause.

Read next

• Tuna fishing goes back 44,000 years: what people have taken from the ray-finned fish lineage, and when.
• Octopus and squid DNA forms flexible loops linked to brain development: another study tracing how a complex trait evolved across a group of species.
• Three-neuron circuit may set the fly’s walking rhythm: behaviour explained through computation, from the inside of one animal.

Sources and disclosure

1. Love J., Nayak A., Grassick A. G., Kong L.-W., Henry E. G., Gil M. A., Fahimipour A. K., Pennell M., Hein A. M. “Evolutionary dynamics of collective behavior across the tree of ray-finned fishes.” PNAS 123 (45): e2611720123 (2026). doi:10.1073/pnas.2611720123. PubMed record: PMID 42993562.
2. Love J. et al. “Collective behavior drives diversification across the tree of ray-finned fishes.” bioRxiv preprint, 8 April 2026. doi:10.64898/2026.04.06.716557.
3. Cornell Chronicle (Krishna Ramanujan), 9 November 2026: Machine learning helps explain evolution of fish schools.
4. Rabosky D. L., Goldberg E. E. “Model inadequacy and mistaken inferences of trait-dependent speciation.” Systematic Biology 64 (2): 340–355 (2015). doi:10.1093/sysbio/syu131.
5. Fahimipour A. K. et al. “Wild animals suppress the spread of socially transmitted misinformation.” PNAS 120 (2023). doi:10.1073/pnas.2215428120.

Disclosure: This article was drafted with AI assistance and checked against the sources above. The full text of the PNAS paper and the preprint were blocked to automated access; SciQuest read the paper’s abstract, author list, affiliations and dates through Crossref and PubMed, the preprint’s abstract and significance statement through Europe PMC and the bioRxiv API, and relied on Cornell University’s release for the gain and loss counts, the methods detail and the quotes. Calculation by SciQuest: 320 ÷ 60 ≈ 5.3 losses per gain. The study was funded by the US National Science Foundation, according to Cornell. The graphic is a SciQuest original; no third-party images are used. Spotted an error? See our corrections policy.